Interpretable Graph Transformer Network for Predicting Adsorption Isotherms of Metal–Organic Frameworks

Abstract

Predicting interactions between metal–organic frameworks (MOFs) and their adsorbates based on structures is critical to design high-performance porous materials. Many gas uptake prediction models have been proposed, but adsorption isotherm prediction is still challenging for most existing models. Here, we report a deep learning approach (MOFNet) that can predict adsorption isotherms for MOFs based on hierarchical representation and pressure adaptive mechanism. We elaborately design a hierarchical representation to encode the MOF structures. We adopt a graph transformer network to capture atomic-level information, which can help learn chemical features required under low-pressure conditions. A pressure adaptive mechanism is employed to interpolate and extrapolate the given limited data points by transfer learning, which can predict adsorption isotherms on a wider pressure range by only one model. We demonstrate that our predictor outperformed other traditional machine learning as well as graph neural network models on the challenging benchmarks and also achieves high performance on the real-world experimental observed adsorption isotherms. Finally, we interpret the models to discover and present potential structure–property relationships using the self-attention mechanism in the network. The proof-of-concept applications, such as disordered MOF predictions and missing data imputation of gas adsorption isotherms, showcase the generality and usability of our model to improve MOF material design.

Coarse-Grained Crystal Graph Neural Networks for Reticular Materials Design

“…Another group of methods includes four crystal graph neural networks: Crystal Graph Convolutional Neural Network (CGCNN), MatErials Graph Network (MEGNet), Global ATtention-based Graph Neural Network with differentiable group normalization and residual connection (DeeperGATGNN), and Atomistic Line Graph Neural Network (ALIGNN). Aside from “general-purpose” structure-aware architectures, a few recent articles introduced neural networks that integrate domain knowledge related to reticular chemistry; MOFNet, MOFTransformer, and MOFormer are worth mentioning. MOF-related models are not a part of our benchmark analysis; the original studies provide an overall picture of accuracy by comparison with crystal graph neural networks, e.g., CGCNN.…”

“…Moreover, the modular structure of reticular materials provides a great opportunity for further tuning of the relevant properties. There were recent efforts to develop specialized featurization schemes for reticular design. − Global descriptors (e.g., topology, volumetric attributes, and energy grids) incorporated into neural network architecture improve the predictive performance and leave room for interpretability analysis. On the other hand, the aforementioned attributes constrain the scenarios where the corresponding models can be applied.…”

A Multi-modal Pre-training Transformer for Universal Transfer Learning in Metal-Organic Frameworks

“…[31][32][33] Due to its superior performance, the Transformer architectures have recently been adopted to predict various properties of MOFs. 34,35 In this work, for the first time in MOF research, we introduce the multi-modal Transformer architecture (named "MOFTransformer"), which captures both the local and global features. Our MOFTransformer was pre-trained with 1 million hypothetical MOFs (hMOFs).…”

Efficient Exploration of Adsorption Space for Separations in Metal–Organic Frameworks Combining the Use of Molecular Simulations, Machine Learning, and Ideal Adsorbed Solution Theory

“…However, these isotherms are strongly focused on a list of small and simple molecules, including H 2 , CH 4 , CO 2 , Xe, Kr, Ar, and N 2 . This focus means that most published ML models predicting single-component or mixture adsorption properties in MOFs are only applicable to a limited collection of adsorbing species. − …”